Rainweic opened a new issue #17164: net.Cast("float16") doesn't work: Check
failed: (*in_type)[i] == dtype_param (2 vs. 0) : This layer requires uniform
type. Expected 'float32' v.s. given 'float16' at 'gamma'
URL: https://github.com/apache/incubator-mxnet/issues/17164
## Description
When I use my dataset to trian Gluoncv's SSD with Float16, I meet this error:
<br>`Check failed: (*in_type)[i] == dtype_param (2 vs. 0) : This layer
requires uniform type. Expected 'float32' v.s. given 'float16' at 'gamma'`<br>
I find in SSD, some BatchNormal Layers is use symbol to achieve. Because
this, `net.cast('float16)` loses efficacy.
### Error Message
mxnet1.5.1:<br>
`Check failed: (*in_type)[i] == dtype_param (2 vs. 0) : This layer requires
uniform type. Expected 'float32' v.s. given 'float16' at 'gamma'`
## To Reproduce
<br>
**Example 1**
```python
import mxnet
import gluoncv
net = gluoncv.model_zoo.get_model('ssd_512_resnet50_v1_voc', \
pretrained=False,
pretrained_base=False,
norm_layer=None,
use_bn=False,
norm_kwargs= None)
net.initialize()
net.cast("float16")
# loses efficacy
one = mxnet.nd.zeros((1,3,512,512), dtype="float16") # meet error
net(one)
```
<br>
**Example 2**
```python
import mxnet as mx
data = mx.sym.var(name="data")
data = mx.sym.Convolution(data, num_filter=512, kernel=(3, 3), pad=(1, 1))
data = mx.sym.Activation(data, act_type="relu")
data = mx.sym.BatchNorm(data)
net = mx.gluon.SymbolBlock(data, mx.sym.var(name='data'))
net.cast("float16")
# loses efficacy
net.initialize()
net(mx.nd.ones((1, 3, 512, 512), dtype='float16')) # meet error
```
### Steps to reproduce
## What have you tried to solve it?
**Example1**
I don't know how to solve it<br>
**Example2**
Change it like this:
```python
import mxnet as mx
data = mx.sym.var(name="data", dtype='float16')
data = mx.sym.Convolution(data, num_filter=512, kernel=(3, 3), pad=(1, 1), )
data = mx.sym.Activation(data, act_type="relu")
data = mx.sym.BatchNorm(data)
net = mx.gluon.SymbolBlock(data, mx.sym.var(name='data', dtype='float16'))
# net.cast("float16")
net.initialize()
net(mx.nd.ones((1, 3, 512, 512), dtype='float16'))
```
## Environment
ubuntu 18 mxnet1.5.1
mac mxnet1.5.1
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
